Difference between DWA local_planner and TEB local_planner

edit

Hi,

the overall idea of both DWA and TEB is to predict/plan the motion of the robot along a given horizon while minimizing a given objective function and while adhering to kinodynamic constraints of the robot. After commanding only the first control action to the robot, the whole prediction/optimization is repeated. This principle is well known in control theory as receding horizon control or model predictive control. However, computing the optimal solution is computationally demanding and thus both approaches approximate the optimal solution in different ways and the actual optimization strategy differs.

Dynamic Window Approach (DWA)

The DWA performs a sample based optimization. It samples a control action in the feasible velocity space (usually a translational/angular velocity pair) and rolls out the trajectory for these particular sampled actions. Rollout means, that the trajectories are simulated according to the specified horizon length based on the robots motion model. One important detail is: the control action is kept constant along the whole prediction horizon. Hence it cannot predict motion reversals etc. After rolling out predictions for all samples, the best candidate is selected based on a specified cost function and constraints (including distance to global path, smoothness, obstacle clearance, ...).

Consequently, DWA includes two simplifications in order to keep the computation times low while achieving a certain amount of control performance. I will not go into the details regarding stability and recursive feasibility arising from the inherent suboptimality, but the approach works pretty well for differential-drive and omnidirectional robots. A benefit is, that the cost function can be non-smooth and so it is well-suited for grid-based evaluations. E.g. trajectories can be rasterized into the costmap in order to evaluate the cost (considering lethal grid cells and inflation cost). Furthermore, the DWA does not get stuck in local minima based on its initialization (of course, the robot can still get stuck due to a limited horizon length and fewer degrees of freedom in terms of control actions). Since DWA assumes constant control actions along the prediction horizon, the control of car-like robots is rather limited. There are some basic extensions to restrict the velocity search space to car-like compatible actions. But motion reversals are still not subject to optimization and hence parking maneuvers in confined spaces are intractable. Note, of course, motion reversals can occure during closed-loop control, but they are not part of the open-loop solution / prediction.

In Summary:

• Suboptimal solutions without motion reversals (control actions are kept constant along the prediction horizon)
• Well-suited for diff-drive/omnidirectional robots, but not for car-like robots
• Supports non-smooth cost functions

Timed-Elastic-Band (TEB)

The TEB primarily tries to seek for the time-optimal solution, but you can also configure it for global reference path fidelity. The approach discretzies the trajectory along the prediction horizon in terms of time and applies a continuous numerical optimization scheme. Consequently, depending on the discretization resolution, the degrees of freedom along the prediciton horizon can be very high and motion reversals are supported. Furthermore, the constrained optimization problem is ... (more)